Source-Aware Spatial Self-Supervision for Sound Event Localization and Detection

This paper presents a source-aware spatial self-supervised learning (SSL) method for sound event localization and detection (SELD) based on neural blind source separation (BSS). SELD involves estimating both the temporal activations of sound classes and their directions of arrival (DOAs) from multichannel audio signals. While deep neural networks have been successful in SELD, their reliance on spatial supervision, including class labels and DOAs, makes annotation challenging and time-intensive. To address this limitation, we propose self-supervised pre-training to separate and localize moving sound sources by maximizing a log-marginal posterior probability of a neural BSS model. This model, called neural full-rank spatial covariance analysis, requires only multichannel mixture signals for training. Our BSS-based framework explicitly trains the model to distinguish multiple sources in a mixture, enabling it to predict DOAs of multiple sound events. The proposed method thus enables us to train a SELD model with a small amount of annotation. The experiments with the STARSS23 dataset show that our method outperforms baseline SELD models trained with a limited amount of supervised data.